Self-driving planners posted much lower errors in offline tests under full-data training with aerial traffic data from a new city, an arXiv preprint reports. On the standard split, RAP's FDE was 2.730 metres under full-data training, compared with 8.116 metres in a zero-shot evaluation using the released checkpoint. DrivoR's FDE was 3.774 metres after full-data training, versus 8.855 metres zero-shot. The paper reports FDE reductions of 66.4% for RAP and 57.4% for DrivoR.
Aerial tracks become local driving examples
The study asks whether a scalable aerial-monitoring alternative can provide learned driving behaviour for adapting autonomous-driving systems to a new city. Its SkyDrive approach treats selected tracked vehicles as virtual egos, then converts their aerial tracks into ego-centric semantic images for trajectory planning and vectorised tracks for motion prediction.
The tracks came from Songdo Traffic monitoring across 20 complex urban intersections over four days, with 10 monitoring time slots per day. The resulting dataset record lists 800 sessions spanning 137.20 hours, along with 367,806 valid vehicles, 234,350 filtered entries and 649,558 ego segments. The train and test portions contain 549,929 and 99,629 segments from 680 and 120 sessions, respectively.
The models learned from a short slice of driving history
For trajectory planning, the models received two seconds of input history and planned the next four seconds at 2Hz. The motion-prediction task used the same two seconds of history but asked models to produce six possible future ego trajectories over six seconds at 10Hz, with probabilities attached.
DrivoR and RAP were evaluated both zero-shot, using author-released checkpoints, and under full-data training. DrivoR was trained from scratch for 20 epochs, while RAP was fine-tuned from the zero-shot checkpoint for five epochs.
Planning errors fell on the standard split
On the standard split, both planners had lower values under full-data training for all four reported measures: ADE, FDE, TTC infraction rate and Non-Compliant Trajectory, or NCT, rate. DrivoR recorded 1.589/3.774/10.82/4.80 for ADE, FDE, TTC and NCT, compared with 3.702/8.855/38.10/8.24 zero-shot. RAP recorded 1.263/2.730/5.89/0.13 under full-data training, compared with 3.399/8.116/25.56/1.12 zero-shot.
The paper describes the differences as substantial reductions in planning error and TTC infractions under in-domain supervision. These are nonrandomized offline comparisons, however, so the reported percentages do not establish that aerial monitoring itself caused the improvement.
The reported gains appeared with limited supervision
The study also tested training subsets of 1%, 5%, 10%, 20% and 50%, corresponding to six, 34, 68, 136 and 340 monitoring sessions. The text reports substantial improvement even with the smallest tested subset, while RAP improved more consistently as the amount of data increased than DrivoR.
Motion prediction showed a sharp local-data gap
Among the motion predictors trained on SongdoDrive, Wayformer had the strongest aggregate results. Its reported BrierFDE was 1.763, minADE was 0.564, minFDE was 1.157 and Miss Rate was 0.148. AutoBot was close behind.
The contrast was larger when a zero-shot MTR* model was compared with trained MTR. MTR* was reported to have 108% higher BrierFDE and 120% higher minADE, with motion-prediction errors also increasing from easy to hard cases.
Wayformer remained strongest across all three difficulty levels. On hard cases, trained MTR achieved a reported 75.8% improvement compared with zero-shot evaluation. MTR* also showed a 341% increase in stationary-type error compared with trained MTR, while turning trajectories were more challenging than straight trajectories.
A promising benchmark result, not a road test
A separate generalization test trained models on 17 intersections and evaluated them on the remaining three. DrivoR's ADE and FDE were 1.485 and 3.514 on the standard split, compared with 1.535 and 3.659 across intersections. RAP's corresponding figures were 1.145 and 2.447, compared with 1.170 and 2.506. The authors describe the gaps as not significant and identify DrivoR's NCT as the clearest gap, although no formal significance test or confidence interval is reported.
The evidence remains limited to offline comparisons in the Songdo setting. It leaves open whether the pattern would carry to other cities, traffic rules, road layouts or driving behaviour, and whether offline gains would appear in closed-loop vehicles or actual collision outcomes.
The document is an arXiv version 1 preprint dated 25 August 2026. The authors state that the data and code will be made publicly available.
Paper data and sources
Original title: SkyDrive: Learning to Drive in a New City from Aerial Traffic Monitoring
Authors: Weijiang Xiong, Lan Feng, Alexandre Alahi, Nikolas Geroliminis
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text